AI TrackerFrontier desk · AEST
Back to news
normalaitech

DecisionRank: On-Policy Distillation for Decision-Model Rerankers

arXiv cs.CL · 31 Oct 2026, 03:56 AEST

DecisionRank introduces a novel approach for training decision models through on-policy distillation. This method achieves state-of-the-art performance on various retrieval benchmarks while significantly improving efficiency.

Key facts

Read aloud

DecisionRank introduces a novel approach for training decision models through on-policy distillation. This method achieves state-of-the-art performance on various retrieval benchmarks while significantly improving efficiency. Decision models target structured tasks by directly predicting choices or scores, avoiding autoregressive decoding latency. To bridge this gap, we introduce DecisionRank, a decision model trained through item-space on-policy distillation OPD. Combined with offline ranking supervision, DecisionRank achieves state-of-the-art performance on both general and reasoning-intensive retrieval benchmarks such as TREC DL, BEIR, and BRIGHT. Even at one.7B parameters, DecisionRank surpasses its 27B generative teacher on TREC DL19--twenty three. At 8B, it outperforms its generative listwise counterpart while running twenty four point five dollars\times$ faster on TREC DL and sixteen point seven dollars\times$ faster on BRIGHT.

Sources in this cluster